The Pre-Con Sandbox: A Practical Guide to Prompting AI for Design-Build Drawing Reviews

In my journey through this industry, I have worn almost every hat. I started out sweeping project sites as a day laborer, swinging a hammer as a carpenter, got my Bachelor’s in Construction Management, managed multi-million-dollar risk as a project engineer on mega-projects and eventually stepped into the world of enterprise software. Across all of those roles, one universal truth has remained constant: 2D drawings are king and the most expensive errors are the ones we build before drawings are checked. 

In a collaborative design-build environment, we don’t have the luxury of waiting for the design to be 100% complete before we start pricing. Design and construction happen concurrently. This speed is our greatest asset, but it also creates a massive coordination challenge. To keep up, many design-build teams are turning to Large Language Models (LLMs) like ChatGPT or Claude to automate drawing reviews.

Yet, if you’ve ever uploaded a sheet and asked an AI, “Is this drawing correct?”, you’ve likely been met with a polite, generic response that completely missed an obvious missing tag. Here’s why that happens and how to run a practical, hands-on experiment that turns any basic AI chatbot into a highly focused pre-con assistant.

The Sandbox Experiment: Auditing a Schedule

To understand how to make AI work for you, we have to look at how it “thinks.” An LLM does not have spatial common sense. It doesn’t know what a concrete wall or an electrical conduit looks like on a PDF. It treats your drawings like a massive, complex puzzle of words, numbers and lines.

If you ask it a broad question, it gets overwhelmed and guesses (resulting in hallucinations). But if you restrict its focus to a simple, mathematical task, like comparing a schedule table to a floor plan, it becomes incredibly effective.

Here’s a practical Structural Prompt you can copy, adapt and use any time you need a stronger starting point.

Imagine you have an architectural set where you want to check that the Door Schedule on Sheet A-601 aligns with the door tags on the Floor Plan on Sheet A-102.

The Copy-and-Paste Prompt Template

Open your LLM of choice, copy the text below and update it for your drawings. Then upload your two drawing sheets (as PDFs or high-resolution images) and run this prompt:

You are a strict, highly detail-oriented pre-construction QA/QC manager on a commercial Design-Build project. Your job is to perform a deterministic audit to find door tags and door schedule discrepancies.

I have uploaded two sheets:
1. Sheet A-601 (containing the Door Schedule table)
2. Sheet A-102 (containing the Floor Plan layout with door tags by door openings)

Perform the following steps with absolute precision:

Step 1: Locate door tags that look like circles (see attached image) on Sheet A-102. Find the door tag associated with this shape and identify its annotated door tag ID (e.g., "100, 101, 102, etc").
Step 2: Go to the Door Schedule on Sheet A-601. Find the row for "Door Tag ID" and verify that that Door Tag from A-102 is or is not present in the Door Schedule.
Step 3: Compare these values. 

If the the Door Schedule includes Door Tags on Sheet A-601 that do not appear on the Floor Plan on Sheet A-102, flag this as a "Door Tag Missing."

Format your output in a clean Markdown table with these exact headers:
| Door Tag | Found on A-102 | Found on A-601 | Status | Notes |

If the values match perfectly, write "No Discrepancy" in the Status column. 
If you cannot find Door Tag 104 on Sheet A-102, do not guess. State "Door Tag 104 not found on floor plan" in the Notes column. Do not write any conversational introduction or conclusion. Provide only the table.

Why This Prompt Works: The Anatomy of Prompt Constraints

If you run this experiment, you will notice the AI suddenly stops giving you vague summaries and starts delivering highly accurate, tabular data. Why?

Let’s break down the mechanics of this prompt using three simple rules:

1. Role: By telling the AI it is a “strict, highly detail-oriented pre-construction QA/QC manager,” you force the algorithm to select words and analytical patterns associated with engineering precision rather than creative writing.
2. Scope: We didn’t ask the AI to “look at the whole set.” We gave it a clear three-step path. We told it exactly which sheets to look at, which room to find and what specific columns to compare. By limiting the search space, we drastically reduce the chance of the AI getting confused or hallucinating.
3. Guardrails: We explicitly told the model: “If you cannot find Room 104, do not guess.” This is crucial. LLMs hate to say “I don’t know.” They are often designed to be helpful, which means they may produce confident-sounding responses even when the better response is to admit it does not have an answer. By giving the machine a designated place to report missing data, we prevent it from making up plausible-sounding results to please you.

Running these manual prompts is a fantastic way to understand the power of AI. It gives your team a sandbox to test ideas, build confidence, and automate small, tedious tasks during pre-construction. 

As a design-build leader, you also have to look at the bigger picture of scale and risk:

• Prompt Scaling: If a single project has 400+ sheets and you’re reviewing 7 projects per month, you cannot have your estimating team manually typing custom prompt scripts and uploading files page-by-page for every unique project. It simply doesn’t scale.
• Data Risk: Public AI tools are trained on user data. Unreleased or proprietary client drawings should be handled carefully; uploading them to a public chat window may raise intellectual property, confidentiality or security risks.

To truly mobilize AI as an enterprise drawing QA/QC tool, we have to move past the chat window. Purpose-built applications like markedup.ai may be a critical part of your drawing audit process because they are designed specifically for this workflow rather than adapted from generalized LLMs. That specialization can help reduce reliance on prompting while better addressing data, accuracy and security considerations.

Use the sandbox to experiment, build your team’s trust and understand the logic of the machine. But when it’s time to protect your GMP and eliminate change orders at scale, transition your workflows from manual prompting to purpose built software.

---